drivers/perf/arm-cmn.c | 121 ++++++++++++++++++++++++++++++----------- 1 file changed, 90 insertions(+), 31 deletions(-)
This series adds support for Graviton5's customised CMN-S3 which has zeroed discovery registers. Patch 1 is based on Robin's suggestion [1] to move the DTM index and watchpoint bitmaps out of hw_perf_event into separate allocations, removing the size constraint that prevented increasing CMN_MAX_DIMENSION. Two fixes were applied on top: the kzalloc size now accounts for sizeof(u64), and bitmap_zero() is used instead of a raw memset for wp_idx. Patch 2 adds the Graviton5 workarounds. [1] https://lore.kernel.org/all/c1d27212-8e2b-4820-bee6-ad7e7de81238@arm.com/ Changes since v3: - Replace patch 1/2 with Robin's approach: dynamically allocate dtm_idx and wp_idx instead of moving the struct into perf_event.h Changes since v2: - Revert DTC logical ID assignment back to xp->logid (per Robin's review) Aviv Bakal (2): perf/arm-cmn: Move DTM index data out of hw_perf_event perf/arm-cmn: Add workarounds for CMN-S3 on Graviton5 drivers/perf/arm-cmn.c | 121 ++++++++++++++++++++++++++++++----------- 1 file changed, 90 insertions(+), 31 deletions(-) -- 2.47.3
This series adds support for Graviton5's customised CMN-S3 which has zeroed discovery registers. Changes since v4: - Fix patch 1 authorship to correctly attribute Robin as author - Collect Reviewed-by tags from Robin and Ilkka Aviv Bakal (1): perf/arm-cmn: Add workarounds for CMN-S3 on Graviton5 Robin Murphy (1): perf/arm-cmn: Move DTM index data out of hw_perf_event drivers/perf/arm-cmn.c | 121 ++++++++++++++++++++++++++++++----------- 1 file changed, 90 insertions(+), 31 deletions(-) -- 2.47.3
From: Robin Murphy <robin.murphy@arm.com>
The amount of data we need to store all the per-DTM counter and
watchpoint allocations is already testing the limits of hw_perf_event,
and future CMNs are only likely to keep growing larger, so move these
arrays out to separate memory allocations. As part of that we can use
an explicit union for allocating cycle counters to dtc_cycles events,
which is arguably nicer anyway.
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Signed-off-by: Aviv Bakal <avivb@amazon.com>
Reviewed-by: Ilkka Koskinen <ilkka@os.amperecomputing.com>
---
drivers/perf/arm-cmn.c | 89 ++++++++++++++++++++++++++++--------------
1 file changed, 59 insertions(+), 30 deletions(-)
diff --git a/drivers/perf/arm-cmn.c b/drivers/perf/arm-cmn.c
index f5305c8fdca4..f1978a53d1c1 100644
--- a/drivers/perf/arm-cmn.c
+++ b/drivers/perf/arm-cmn.c
@@ -597,17 +597,14 @@ static void arm_cmn_debugfs_init(struct arm_cmn *cmn, int id) {}
struct arm_cmn_hw_event {
struct arm_cmn_node *dn;
- u64 dtm_idx[DIV_ROUND_UP(CMN_MAX_NODES_PER_EVENT * 2, 64)];
+ union {
+ u64 *dtm_idx;
+ int cc_idx;
+ };
+ unsigned long *wp_idx;
s8 dtc_idx[CMN_MAX_DTCS];
u8 num_dns;
u8 dtm_offset;
-
- /*
- * WP config registers are divided to UP and DOWN events. We need to
- * keep to track only one of them.
- */
- DECLARE_BITMAP(wp_idx, CMN_MAX_XPS);
-
bool wide_sel;
enum cmn_filter_select filter_sel;
};
@@ -625,25 +622,42 @@ static struct arm_cmn_hw_event *to_cmn_hw(struct perf_event *event)
return (struct arm_cmn_hw_event *)&event->hw;
}
-static void arm_cmn_set_index(u64 x[], unsigned int pos, unsigned int val)
+static void arm_cmn_set_dtm_idx(struct arm_cmn_hw_event *hw, unsigned int pos, unsigned int val)
{
- x[pos / 32] |= (u64)val << ((pos % 32) * 2);
+ hw->dtm_idx[pos / 32] |= (u64)val << ((pos % 32) * 2);
}
-static unsigned int arm_cmn_get_index(u64 x[], unsigned int pos)
+static unsigned int arm_cmn_get_dtm_idx(struct arm_cmn_hw_event *hw, unsigned int pos)
{
- return (x[pos / 32] >> ((pos % 32) * 2)) & 3;
+ return (hw->dtm_idx[pos / 32] >> ((pos % 32) * 2)) & 3;
}
-static void arm_cmn_set_wp_idx(unsigned long *wp_idx, unsigned int pos, bool val)
+static u64 *arm_cmn_alloc_dtm_idx(void)
+{
+ return kzalloc(DIV_ROUND_UP(CMN_MAX_NODES_PER_EVENT * 2, 64) * sizeof(u64), GFP_KERNEL);
+}
+
+static void arm_cmn_set_wp_idx(struct arm_cmn_hw_event *hw, unsigned int pos, bool val)
{
if (val)
- set_bit(pos, wp_idx);
+ set_bit(pos, hw->wp_idx);
+}
+
+static unsigned int arm_cmn_get_wp_idx(struct arm_cmn_hw_event *hw, unsigned int pos)
+{
+ return test_bit(pos, hw->wp_idx);
+}
+
+static unsigned long *arm_cmn_alloc_wp_idx(void)
+{
+ return bitmap_zalloc(CMN_MAX_XPS, GFP_KERNEL);
}
-static unsigned int arm_cmn_get_wp_idx(unsigned long *wp_idx, unsigned int pos)
+static void arm_cmn_clear_idx(struct arm_cmn_hw_event *hw)
{
- return test_bit(pos, wp_idx);
+ memset(hw->dtm_idx, 0, DIV_ROUND_UP(CMN_MAX_NODES_PER_EVENT * 2, 64) * sizeof(u64));
+ if (hw->wp_idx)
+ bitmap_zero(hw->wp_idx, CMN_MAX_XPS);
}
struct arm_cmn_event_attr {
@@ -1376,7 +1390,7 @@ static int arm_cmn_get_assigned_wp_idx(struct perf_event *event,
struct arm_cmn_hw_event *hw,
unsigned int pos)
{
- return CMN_EVENT_EVENTID(event) + arm_cmn_get_wp_idx(hw->wp_idx, pos);
+ return CMN_EVENT_EVENTID(event) + arm_cmn_get_wp_idx(hw, pos);
}
static void arm_cmn_claim_wp_idx(struct arm_cmn_dtm *dtm,
@@ -1387,7 +1401,7 @@ static void arm_cmn_claim_wp_idx(struct arm_cmn_dtm *dtm,
struct arm_cmn_hw_event *hw = to_cmn_hw(event);
dtm->wp_event[wp_idx] = hw->dtc_idx[dtc];
- arm_cmn_set_wp_idx(hw->wp_idx, pos, wp_idx - CMN_EVENT_EVENTID(event));
+ arm_cmn_set_wp_idx(hw, pos, wp_idx - CMN_EVENT_EVENTID(event));
}
static u32 arm_cmn_wp_config(struct perf_event *event, int wp_idx)
@@ -1458,7 +1472,7 @@ static u64 arm_cmn_read_dtm(struct arm_cmn *cmn, struct arm_cmn_hw_event *hw,
dtm = &cmn->dtms[dn->dtm] + hw->dtm_offset;
reg = readq_relaxed(dtm->base + offset);
}
- dtm_idx = arm_cmn_get_index(hw->dtm_idx, i);
+ dtm_idx = arm_cmn_get_dtm_idx(hw, i);
count += (u16)(reg >> (dtm_idx * 16));
}
return count;
@@ -1505,7 +1519,7 @@ static void arm_cmn_event_read(struct perf_event *event)
unsigned long flags;
if (CMN_EVENT_TYPE(event) == CMN_TYPE_DTC) {
- delta = arm_cmn_read_cc(cmn->dtc + hw->dtc_idx[0]);
+ delta = arm_cmn_read_cc(cmn->dtc + hw->cc_idx);
local64_add(delta, &event->count);
return;
}
@@ -1572,7 +1586,7 @@ static void arm_cmn_event_start(struct perf_event *event, int flags)
int i;
if (type == CMN_TYPE_DTC) {
- struct arm_cmn_dtc *dtc = cmn->dtc + hw->dtc_idx[0];
+ struct arm_cmn_dtc *dtc = cmn->dtc + hw->cc_idx;
writel_relaxed(CMN_DT_DTC_CTL_DT_EN | CMN_DT_DTC_CTL_CG_DISABLE,
dtc->base + CMN_DT_DTC_CTL);
@@ -1590,7 +1604,7 @@ static void arm_cmn_event_start(struct perf_event *event, int flags)
writeq_relaxed(mask, base + CMN_DTM_WPn_MASK(wp_idx));
}
} else for_each_hw_dn(hw, dn, i) {
- int dtm_idx = arm_cmn_get_index(hw->dtm_idx, i);
+ int dtm_idx = arm_cmn_get_dtm_idx(hw, i);
arm_cmn_set_event_sel_lo(dn, dtm_idx, CMN_EVENT_EVENTID(event),
hw->wide_sel);
@@ -1606,7 +1620,7 @@ static void arm_cmn_event_stop(struct perf_event *event, int flags)
int i;
if (type == CMN_TYPE_DTC) {
- struct arm_cmn_dtc *dtc = cmn->dtc + hw->dtc_idx[0];
+ struct arm_cmn_dtc *dtc = cmn->dtc + hw->cc_idx;
dtc->cc_active = false;
writel_relaxed(CMN_DT_DTC_CTL_DT_EN, dtc->base + CMN_DT_DTC_CTL);
@@ -1619,7 +1633,7 @@ static void arm_cmn_event_stop(struct perf_event *event, int flags)
writeq_relaxed(~0ULL, base + CMN_DTM_WPn_VAL(wp_idx));
}
} else for_each_hw_dn(hw, dn, i) {
- int dtm_idx = arm_cmn_get_index(hw->dtm_idx, i);
+ int dtm_idx = arm_cmn_get_dtm_idx(hw, i);
arm_cmn_set_event_sel_lo(dn, dtm_idx, 0, hw->wide_sel);
}
@@ -1763,6 +1777,14 @@ static enum cmn_filter_select arm_cmn_filter_sel(const struct arm_cmn *cmn,
}
+static void arm_cmn_event_destroy(struct perf_event *event)
+{
+ struct arm_cmn_hw_event *hw = to_cmn_hw(event);
+
+ kfree(hw->dtm_idx);
+ bitmap_free(hw->wp_idx);
+}
+
static int arm_cmn_event_init(struct perf_event *event)
{
struct arm_cmn *cmn = to_cmn(event->pmu);
@@ -1787,6 +1809,11 @@ static int arm_cmn_event_init(struct perf_event *event)
if (type == CMN_TYPE_DTC)
return arm_cmn_validate_group(cmn, event);
+ event->destroy = arm_cmn_event_destroy;
+ hw->dtm_idx = arm_cmn_alloc_dtm_idx();
+ if (!hw->dtm_idx)
+ return -ENOMEM;
+
eventid = CMN_EVENT_EVENTID(event);
/* For watchpoints we need the actual XP node here */
if (type == CMN_TYPE_WP) {
@@ -1797,6 +1824,9 @@ static int arm_cmn_event_init(struct perf_event *event)
/* ...but the DTM may depend on which port we're watching */
if (cmn->multi_dtm)
hw->dtm_offset = CMN_EVENT_WP_DEV_SEL(event) / 2;
+ hw->wp_idx = arm_cmn_alloc_wp_idx();
+ if (!hw->wp_idx)
+ return -ENOMEM;
} else if (type == CMN_TYPE_XP &&
(cmn->part == PART_CMN700 || cmn->part == PART_CMN_S3)) {
hw->wide_sel = true;
@@ -1847,7 +1877,7 @@ static void arm_cmn_event_clear(struct arm_cmn *cmn, struct perf_event *event,
while (i--) {
struct arm_cmn_dtm *dtm = &cmn->dtms[hw->dn[i].dtm] + hw->dtm_offset;
- unsigned int dtm_idx = arm_cmn_get_index(hw->dtm_idx, i);
+ unsigned int dtm_idx = arm_cmn_get_dtm_idx(hw, i);
if (type == CMN_TYPE_WP) {
int wp_idx = arm_cmn_get_assigned_wp_idx(event, hw, i);
@@ -1861,8 +1891,7 @@ static void arm_cmn_event_clear(struct arm_cmn *cmn, struct perf_event *event,
dtm->pmu_config_low &= ~CMN__PMEVCNT_PAIRED(dtm_idx);
writel_relaxed(dtm->pmu_config_low, dtm->base + CMN_DTM_PMU_CONFIG);
}
- memset(hw->dtm_idx, 0, sizeof(hw->dtm_idx));
- memset(hw->wp_idx, 0, sizeof(hw->wp_idx));
+ arm_cmn_clear_idx(hw);
for_each_hw_dtc_idx(hw, j, idx)
cmn->dtc[j].counters[idx] = NULL;
@@ -1882,7 +1911,7 @@ static int arm_cmn_event_add(struct perf_event *event, int flags)
return -ENOSPC;
cmn->dtc[i].cycles = event;
- hw->dtc_idx[0] = i;
+ hw->cc_idx = i;
if (flags & PERF_EF_START)
arm_cmn_event_start(event, 0);
@@ -1947,7 +1976,7 @@ static int arm_cmn_event_add(struct perf_event *event, int flags)
goto free_dtms;
}
- arm_cmn_set_index(hw->dtm_idx, i, dtm_idx);
+ arm_cmn_set_dtm_idx(hw, i, dtm_idx);
dtm->input_sel[dtm_idx] = input_sel;
shift = CMN__PMEVCNTn_GLOBAL_NUM_SHIFT(dtm_idx);
@@ -1980,7 +2009,7 @@ static void arm_cmn_event_del(struct perf_event *event, int flags)
arm_cmn_event_stop(event, PERF_EF_UPDATE);
if (type == CMN_TYPE_DTC)
- cmn->dtc[hw->dtc_idx[0]].cycles = NULL;
+ cmn->dtc[hw->cc_idx].cycles = NULL;
else
arm_cmn_event_clear(cmn, event, hw->num_dns);
}
--
2.47.3
Graviton5 uses a customised CMN-S3 implementation where certain
discovery registers report zeroed fields. Add the following workarounds:
- Introduce a dedicated ACPI HID to identify the Graviton5 CMN variant.
- Derive the DTC domain from the XP node ID, since the unit info
register reports it as zero.
- Set the DTC logical ID from the XP's logical ID, since the node info
register's logical ID field is also zeroed.
Signed-off-by: Aviv Bakal <avivb@amazon.com>
Reviewed-by: Robin Murphy <robin.murphy@arm.com>
Reviewed-by: Ilkka Koskinen <ilkka@os.amperecomputing.com>
---
drivers/perf/arm-cmn.c | 32 +++++++++++++++++++++++++++++++-
1 file changed, 31 insertions(+), 1 deletion(-)
diff --git a/drivers/perf/arm-cmn.c b/drivers/perf/arm-cmn.c
index f1978a53d1c1..3fb71d9a57eb 100644
--- a/drivers/perf/arm-cmn.c
+++ b/drivers/perf/arm-cmn.c
@@ -31,7 +31,8 @@
#define CMN_CHILD_NODE_ADDR GENMASK(29, 0)
#define CMN_CHILD_NODE_EXTERNAL BIT(31)
-#define CMN_MAX_DIMENSION 12
+/* Some implementations use a mesh larger than the architectural max of 12 */
+#define CMN_MAX_DIMENSION 14
#define CMN_MAX_XPS (CMN_MAX_DIMENSION * CMN_MAX_DIMENSION)
#define CMN_MAX_DTMS (CMN_MAX_XPS + (CMN_MAX_DIMENSION - 1) * 4)
@@ -214,6 +215,8 @@ enum cmn_part {
PART_CMN700 = 0x43c,
PART_CI700 = 0x43a,
PART_CMN_S3 = 0x43e,
+ /* Synthetic part number, overridden to PART_CMN_S3 during discovery */
+ PART_GRAVITON5 = 0xa5,
};
/* CMN-600 r0px shouldn't exist in silicon, thankfully */
@@ -2250,6 +2253,18 @@ static unsigned int arm_cmn_dtc_domain(struct arm_cmn *cmn, void __iomem *xp_reg
return FIELD_GET(CMN_DTM_UNIT_INFO_DTC_DOMAIN, readl_relaxed(xp_region + offset));
}
+static unsigned int arm_cmn_graviton5_dtc_domain(u16 xp_id)
+{
+ unsigned int x = (xp_id >> 7) & 0xf;
+ unsigned int y = (xp_id >> 3) & 0xf;
+
+ /*
+ * The unit info register reads as zero; derive the DTC domain from
+ * the XP's mesh coordinates over the 10x14 mesh.
+ */
+ return (x / 5) + (y / 7) * 2;
+}
+
static void arm_cmn_init_node_info(struct arm_cmn *cmn, u32 offset, struct arm_cmn_node *node)
{
int level;
@@ -2295,6 +2310,7 @@ static int arm_cmn_discover(struct arm_cmn *cmn, unsigned int rgn_offset)
u64 reg;
int i, j;
size_t sz;
+ bool graviton5_workaround = false;
arm_cmn_init_node_info(cmn, rgn_offset, &cfg);
if (cfg.type != CMN_TYPE_CFG)
@@ -2305,6 +2321,13 @@ static int arm_cmn_discover(struct arm_cmn *cmn, unsigned int rgn_offset)
reg = readq_relaxed(cfg_region + CMN_CFGM_PERIPH_ID_01);
part = FIELD_GET(CMN_CFGM_PID0_PART_0, reg);
part |= FIELD_GET(CMN_CFGM_PID1_PART_1, reg) << 8;
+
+ /* Graviton5 has a customised CMN-S3 which needs some fixups */
+ if (cmn->part == PART_GRAVITON5) {
+ cmn->part = PART_CMN_S3;
+ graviton5_workaround = true;
+ }
+
/* 600AE is close enough that it's not really worth more complexity */
if (part == PART_CMN600AE)
part = PART_CMN600;
@@ -2394,6 +2417,8 @@ static int arm_cmn_discover(struct arm_cmn *cmn, unsigned int rgn_offset)
if (cmn->part == PART_CMN600)
xp->dtc = -1;
+ else if (graviton5_workaround)
+ xp->dtc = arm_cmn_graviton5_dtc_domain(xp->id);
else
xp->dtc = arm_cmn_dtc_domain(cmn, xp_region);
@@ -2472,6 +2497,10 @@ static int arm_cmn_discover(struct arm_cmn *cmn, unsigned int rgn_offset)
switch (dn->type) {
case CMN_TYPE_DTC:
+ if (graviton5_workaround) {
+ /* Node info logical ID is zeroed; use the XP's */
+ dn->logid = xp->logid;
+ }
cmn->num_dtcs++;
dn++;
break;
@@ -2687,6 +2716,7 @@ static const struct acpi_device_id arm_cmn_acpi_match[] = {
{ "ARMHC650" },
{ "ARMHC700" },
{ "ARMHC003" },
+ { "AMZN0070", PART_GRAVITON5 },
{}
};
MODULE_DEVICE_TABLE(acpi, arm_cmn_acpi_match);
--
2.47.3
© 2016 - 2026 Red Hat, Inc.